ConvReLU3d

Fused float Conv3d + ReLU marker — a plain-float lucid.nn.Sequential.

A thin lucid.nn.Sequential of [conv, relu] emitted by lucid.quantization.fuse_modules. It runs as plain float — it carries no observers and no fake-quant, and computing through it is identical to running the conv then the ReLU. Its whole purpose is structural: by packaging the two ops as one unit it tells the quantization pipeline to place a single activation observer after the ReLU rather than between the conv and the ReLU.

Why fuse for quantization. If an observer sat on the raw conv output it would calibrate a symmetric range that includes values the ReLU is about to discard, wasting half the int8 codes on negatives that never reach inference. Observing the post-ReLU range instead lets the eventual quantized layer pick a grid over $[0, \max]$, roughly doubling resolution. Fusion also drops the intermediate float tensor between conv and ReLU, so the converted kernel does conv → ReLU → requantize in one step.

Where it sits. This marker is the hand-off point between float and quantized worlds. lucid.quantization.prepare_qat replaces it with a trainable lucid.nn.qat.ConvReLU3d (weight + post-ReLU fake-quant under an STE), and lucid.quantization.convert maps it — or the QAT layer it became — to a fused inference lucid.nn.quantized.ConvReLU3d.

As a plain-float module it simply computes, with $\star$ the 3-D cross-correlation:

Downstream, once an observer is attached, that fused output is fake-quantized under a straight-through estimator — rounded in the forward pass, identity in the backward: